Isa Yildirim

Approximate spectral clustering with utilized similarity information using geodesic based hybrid distance measures

Spectral clustering has been popular thanks to its ability to extract clusters of varying characteristics without using a parametric model in expense of high computational cost required for eigendecomposition of pairwise similarities. In order to utilize its advantages in large datasets where it is infeasible due to its computational burden, approximate spectral clustering (ASC) methods apply spectral clustering on a reduced set of points (data representatives) selected by sampling or quantization. This two-step approach (i.e. finding the representatives and then clustering them) brings new opportunities for precise similarity definition such as manifold based topological relations, data distribution within the Voronoi polyhedra of the representatives, and their geodesic distance information, which are often ignored in similarity definition for ASC. In this study, we propose geodesic based hybrid similarity criteria which enable the use of different types of information for accurate similarity representation in ASC. Despite the fact that geodesic concept has been widely used in clustering, our contribution is the unique way of representing data topology to form geodesic relations and jointly harnessing various information types including topology, distance and density. The proposed criteria are tested using both sampling (selective sampling) and quantization (neural gas and k-means++) approaches. Experiments on artificial datasets, well-known small/medium-size real datasets, and four large datasets (four remote-sensing images), with different types of clusters, show that the proposed geodesic based hybrid similarity criteria outperform traditional similarity criteria in terms of clustering accuracies and several cluster validity indices. HighlightsGeodesic based hybrid similarity measures are proposed for approximate spectral clustering.Neighborhood graph, required for geodesic approach, is effectively constructed by weighted Delaunay triangulation (CONN).The proposed geodesic based hybrid similarities outperform in terms of both accuracies and cluster validity indices.The proposed geodesic based hybrid similarities can be powerful for clustering of large remote sensing images.The proposed similarities are significant especially for quantization based approximate spectral clustering.

Digital breast tomosynthesis image reconstruction using 2D and 3D total variation minimization

BackgroundDigital breast tomosynthesis (DBT) is an emerging imaging modality which produces three-dimensional radiographic images of breast. DBT reconstructs tomographic images from a limited view angle, thus data acquired from DBT is not sufficient enough to reconstruct an exact image. It was proven that a sparse image from a highly undersampled data can be reconstructed via compressed sensing (CS) techniques. This can be done by minimizing the l1 norm of the gradient of the image which can also be defined as total variation (TV) minimization. In tomosynthesis imaging problem, this idea was utilized by minimizing total variation of image reconstructed by algebraic reconstruction technique (ART). Previous studies have largely addressed 2-dimensional (2D) TV minimization and only few of them have mentioned 3-dimensional (3D) TV minimization. However, quantitative analysis of 2D and 3D TV minimization with ART in DBT imaging has not been studied.MethodsIn this paper two different DBT image reconstruction algorithms with total variation minimization have been developed and a comprehensive quantitative analysis of these two methods and ART has been carried out: The first method is ART + TV2D where TV is applied to each slice independently. The other method is ART + TV3D in which TV is applied by formulating the minimization problem 3D considering all slices.ResultsA 3D phantom which roughly simulates a breast tomosynthesis image was designed to evaluate the performance of the methods both quantitatively and qualitatively in the sense of visual assessment, structural similarity (SSIM), root means square error (RMSE) of a specific layer of interest (LOI) and total error values. Both methods show superior results in reducing out-of-focus slice blur compared to ART.ConclusionsComputer simulations show that ART + TV3D method substantially enhances the reconstructed image with fewer artifacts and smaller error rates than the other two algorithms under the same configuration and parameters and it provides faster convergence rate.

An iterative tomosynthesis reconstruction using total variation combined with non-local means filtering

BackgroundAfter the release of compressed sensing (CS) theory, reconstruction algorithms from sparse and incomplete data have shown great improvements in diminishing artifacts of missing data. Following this progress, both local and non-local regularization induced iterative reconstructions have been actively used in limited view angle imaging problems.MethodsIn this study, a 3D iterative image reconstruction method (ART + TV)NLM was introduced by combining local total variation (TV) with non-local means (NLM) filter. In the first step, TV minimization was applied to the image obtained by algebraic reconstruction technique (ART) for background noise removal with preserving edges. In the second step, NLM is used in order to suppress the out of focus slice blur which is the most existent image artifact in tomosynthesis imaging. NLM exploits the similar structures to increase the smoothness in the image reconstructed by ART + TV.ResultsA tomosynthesis system and a 3D phantom were designed to perform simulations to show the superior performance of our proposed (ART + TV)NLM over ART and widely used ART + TV methods. Visual inspections show a significant improvement in image quality compared to ART and ART + TV.ConclusionsRMSE, Structure SIMilarity (SSIM) value and SNR of a specific layer of interest (LOI) showed that by proper selection of NLM parameters, significant improvements can be achieved in terms of convergence rate and image quality.

An Approximate Spectral Clustering Ensemble for High Spatial Resolution Remote-Sensing Images

Unsupervised clustering of high spatial resolution remote-sensing images plays a significant role in detailed land-cover identification, especially for agricultural and environmental monitoring. A recently promising method is approximate spectral clustering (SC) which enables spectral partitioning for large datasets to extract clusters with distinct characteristics without a parametric model. It also facilitates the use of various information types via advanced similarity criteria. However, it requires an empirical selection of a similarity criterion optimal for the corresponding application. To address this challenge, we propose an approximate SC ensemble (ASCE2) which fuses partitionings obtained by different similarity representations. Contrary to existing spectral ensembles for remote-sensing applications, the proposed ASCE2 employs neural gas quantization instead of random sampling, advanced similarity criteria instead of traditional distance-based Gaussian kernel with different decay parameters, and a two-level ensemble. We evaluate the proposed ASCE2 with three measures (accuracy, adjusted Rand index, and normalized mutual information) using five remote-sensing images, two of which are commonly available. We apply the ASCE2 in two applications for agricultural monitoring: 1) land-cover identification to determine orchard fields using a WorldView-2 image (0.5-m spatial resolution) and 2) finding lands in good agricultural condition using multitemporal RapidEye images (5-m spatial resolution). Experimental results indicate a significant betterment of the resulting partitionings obtained by the proposed ensemble, with respect to the evaluation measures in these applications.

Geodesic Based Similarities for Approximate Spectral Clustering

Spectral clustering has been successfully used in various applications, thanks to its properties such as no requirement of a parametric model, ability to extract clusters of different characteristics and easy implementation. However, it is often infeasible for large datasets due to its heavy computational load and memory requirement. To utilize its advantages for large datasets, it is applied to the dataset representatives (either obtained by quantization or sampling) rather than the data samples, which is called approximate spectral clustering. This necessitates novel approaches for defining similarities based on representatives exploiting the data characteristics, in addition to the traditional Euclidean distance based similarities. To address this challenge, we propose similarity measures based on geodesic distances and local density distribution. Our experiments using datasets with varying cluster statistics show that the proposed geodesic based similarities are successful for approximate spectral clustering with high accuracies.

Iterative image reconstruction using non-local means with total variation from insufficient projection data

In this work, algebraic reconstruction technique (ART) is extended by using non-local means (NLM) and total variation (TV) for reduction of artifacts that are due to insufficient projection data. TV and NLM algorithms use different image models and their application in tandem becomes a powerful denoising method that reduces erroneous variations in the image while preserving edges and details. Simulations were performed on a widely used 2D Shepp-Logan phantom to demonstrate performance of the introduced method (ART + TV) NLM and compare it to TV based ART (ART + TV) and ART. The results indicate that (ART + TV) NLM achieves better reconstructions compared to (ART + TV) and ART.

Automated respiratory phase and onset detection using only chest sound signal

The problem of non-invasive detection of respiratory phases and onsets without making direct airflow measurement is addressed here. Currently available techniques require the use of multichannel recorded sounds of both chest and trachea. In this paper, we propose a method which detects both respiratory phases and onsets using only chest sound data. Prior signal information in both time and frequency from the chest sound is exploited to isolate the lung component of the sound and the quasi-periodicity of its short-term energy is used to develop a configuration of nonlinear filters and bandpass filters to estimate the respiratory phase onsets. Performance results for the proposed method are reported for the case of low and medium flow rates. The average onset localizing accuracy of the proposed method is shown to be comparable to that obtained with data from more than one recording channel.

A Robust Method to Estimate Time Split in Second Heart Sound Using Instantaneous Frequency Analysis

Closure of the aortic valve (A2) and the pulmonary valve (P2) generates the second heart sound (S2). The time separation between A2 and P2 is known as the A2- P2 split and it has very important diagnostic potential. Methods proposed in the past to measure the split noninvasively are limited either by prior signal modeling assumptions or by reliance on manual processing in key steps. In this work, we propose a new method that is devised to noninvasively provide an automated measurement of the time split between A2 and P2 with minimal prior assumptions on signal models. Our method is based on tracking the changes of the instantaneous frequency (IF) of S2 via time frequency representation of the S2 obtained by smoothed Wigner-Ville Distribution. The cues provided by the changes in the IF trajectory are analyzed using an automated procedure to identify the onset of the P2 pulse. Simulations are carried out to demonstrate the effectiveness of the procedure in estimating the split. The performance of the method in the presence of noise varying between 6 dB and 8 dB for several trials and interference is investigated and the robustness of the method is demonstrated.

Automated nanofiber diameter measurement in SEM images using a robust image analysis method

Due to the high surface area, porosity, and rigidity, applications of nanofibers and nanosurfaces have developed in recent years. Nanofibers and nanosurfaces are typically produced by electrospinning method. In the production process, determination of average fiber diameter is crucial for quality assessment. Average fiber diameter is determined by manually measuring the diameters of randomly selected fibers on scanning electron microscopy (SEM) images. However, as the number of the images increases, manual fiber diameter determination becomes a tedious and time consuming task as well as being sensitive to human errors. Therefore, an automated fiber diameter measurement system is desired. In the literature, this task is achieved by using image analysis algorithms. Typically, these methods first isolate each fiber in the image and measure the diameter of each isolated fiber. Fiber isolation is an error-prone process. In this study, automated calculation of nanofiber diameter is achieved without fiber isolation using image processing and analysis algorithms. Performance of the proposed method was tested on real data. The effectiveness of the proposed method is shown by comparing automatically and manually measured nanofiber diameter values.

Regularized Estimation of Retinal Vascular Oxygen Tension From Phosphorescence Images

The level of retinal oxygenation is potentially an important cue to the onset or presence of some common retinal diseases. An improved method for assessing oxygen tension in retinal blood vessels from phosphorescence lifetime imaging data is reported in this paper. The optimum estimate for phosphorescence lifetime and oxygen tension is obtained by regularizing the least-squares (LS) method. The estimation method is implemented with an iterative algorithm to minimize a regularized LS cost function. The effectiveness of the proposed method is demonstrated by applying it to simulated data as well as image data acquired from rat retinas. The method is shown to yield estimates that are robust to noise and whose variance is lower than that obtained with the classical LS method.